A value of type Backend represents one of GHC's back ends. The set of back ends cannot be extended except by modifying the definition of Backend in this module.

The Backend type is abstract; that is, its value constructors are not exported. It's crucial that they not be exported, because a value of type Backend carries only the back end's name, not its behavior or properties. If Backend were not abstract, then code elsewhere in the compiler could depend directly on the name, not on the semantics, which would make it challenging to create a new back end. Because Backend is abstract, all the obligations of a new back end are enumerated in this module, in the form of functions that take Backend as an argument.

The issue of abstraction is discussed at great length in #20927 and !7442.

The native code generator. Compiles Cmm code into textual assembler, then relies on an external assembler toolchain to produce machine code.

Only supports a few platforms (X86, PowerPC, SPARC).

See GHC.CmmToAsm.

The LLVM backend.

Compiles Cmm code into LLVM textual IR, then relies on LLVM toolchain to produce machine code.

It relies on LLVM support for the calling convention used by the NCG backend to produce code objects ABI compatible with it (see "cc 10" or "ghccc" calling convention in https://llvm.org/docs/LangRef.html#calling-conventions).

Supports a few platforms (X86, AArch64, s390x, ARM).

See GHC.CmmToLlvm

The JavaScript Backend

See documentation in GHC.StgToJS

Via-C ("unregisterised") backend.

Compiles Cmm code into C code, then relies on a C compiler to produce machine code.

It produces code objects that are not ABI compatible with those produced by NCG and LLVM backends.

Produced code is expected to be less efficient than the one produced by NCG and LLVM backends because STG registers are not pinned into real registers. On the other hand, it supports more target platforms (those having a valid C toolchain).

See GHC.CmmToC

The ByteCode interpreter.

Produce ByteCode objects (BCO, see GHC.ByteCode) that can be interpreted. It is used by GHCi.

Currently some extensions are not supported (foreign primops).

See GHC.StgToByteCode

A dummy back end that generates no code.

Use this back end to disable code generation. It is particularly useful when GHC is used as a library for other purpose than generating code (e.g. to generate documentation with Haddock) or when the user requested it (via `-fno-code`) for some reason.

A list of all back ends. They are ordered as we wish them to appear when they are enumerated in error messages.

This enumeration type specifies how the back end wishes GHC's primitives to be implemented. (Module GHC.StgToCmm.Prim provides a generic implementation of every primitive, but some primitives, like IntQuotRemOp, can be implemented more efficiently by certain back ends on certain platforms. For example, by using a machine instruction that simultaneously computes quotient and remainder.)

For the meaning of each alternative, consult GHC.StgToCmm.Config. (In a perfect world, type PrimitiveImplementation would be defined there, in the module that determines its meaning. But I could not figure out how to do it without mutual recursion across module boundaries.)

Names a function that generates code and writes the results to a file, of this type:

   Logger
-> DynFlags
-> Module -- ^ module being compiled
-> ModLocation
-> FilePath -- ^ Where to write output
-> Set UnitId -- ^ dependencies
-> Stream IO RawCmmGroup a -- results from `StgToCmm`
-> IO a

The functions so named are defined in GHC.Driver.CodeOutput.

We expect one function per back end—or more precisely, one function for each back end that writes code to a file. (The interpreter does not write to files; its output lives only in memory.)

Names a function that tells the driver what should happen after assembly code is written. This might include running a C compiler, running LLVM, running an assembler, or various similar activities. The function named normally has this type:

   TPipelineClass TPhase m
=> PipeEnv
-> HscEnv
-> Maybe ModLocation
-> FilePath
-> m (Maybe FilePath)

The functions so named are defined in GHC.Driver.Pipeline.

Names a function that tells the driver what command-line options to include when invoking a C compiler. It's meant for -D options that define symbols for the C preprocessor. Because the exact symbols defined might depend on versions of tools located in the file system (cough LLVM cough), the function requires an IO action. The function named has this type:

Logger -> DynFlags -> IO [String]

An informal description of the back end, for use in issuing warning messages only. If code depends on what's in the string, you deserve what happens to you.

This flag tells the compiler driver whether the back end will write files: interface files and object files. It is typically true for "real" back ends that generate code into the filesystem. (That means, not the interpreter.)

When the back end does write files, this value tells the compiler in what manner of file the output should go: temporary, persistent, or specific.

This flag tells the driver whether the back end can reuse code (bytecode or object code) that has been loaded dynamically. Likely true only of the interpreter.

It is is true of every back end except -fno-code that it "generates code." Surprisingly, this property influences the driver in a ton of ways. Some examples:

• If the back end does not generate code, then the driver needs to turn on code generation for Template Haskell (because that code needs to be generated and run at compile time).
• If the back end does not generate code, then the driver does not need to deal with an output file.
• If the back end does generated code, then the driver supports HscRecomp. If not, recompilation does not need a linkable (and is automatically up to date).

When set, this flag turns on interface writing for Backpack. It should probably be the same as backendGeneratesCode, but it is kept distinct for reasons described in Note [-fno-code mode].

When preparing code for this back end, the type checker should pay attention to SPECIALISE pragmas. If this flag is False, then the type checker ignores SPECIALISE pragmas (for imported things?).

This back end wants the mi_globals field of a ModIface to be populated (with the top-level bindings of the original source). Only true for the interpreter.

The back end targets a technology that implements switch natively. (For example, LLVM or C.) Therefore it is not necessary for GHC to ccompile a Cmm Switch form into a decision tree with jump tables at the leaves.

As noted in the documentation for PrimitiveImplementation, certain primitives have multiple implementations, depending on the capabilities of the back end. This field signals to module GHC.StgToCmm.Prim what implementations to use with this back end.

When this value is IsValid, the back end is compatible with vector instructions. When it is NotValid, it carries a message that is shown to users.

This flag says whether the back end supports large binary blobs. See Note [Embedding large binary blobs] in GHC.CmmToAsm.Ppr.

This flag tells the compiler driver that the back end does not support every target platform; it supports only platforms that claim NCG support. (It's set only for the native code generator.) Crufty. If the driver tries to use the native code generator without platform support, the driver fails over to the LLVM back end.

This flag is set if the back end can generate code for proc points. If the flag is not set, then a Cmm pass needs to split proc points (that is, turn each proc point into a standalone procedure).

This flag guides the driver in resolving issues about API support on the target platform. If the flag is set, then these things are true:

• When the target platform supports only an unregisterised API, this backend can be replaced with compilation via C.
• When the target does not support an unregisterised API, this back end can replace compilation via C.

This flag is true if the back end works *only* with the unregisterised ABI.

This flag is set if the back end generates C code in a .hc file. The flag lets the compiler driver know if the command-line flag -C is meaningful.

This flag says whether SPT (static pointer table) entries will be inserted dynamically if needed. If this flag is False, then GHC.Iface.Tidy should emit C stubs that initialize the SPT entries.

If this flag is unset, then the driver ignores the flag -fbreak-points, since backends other than the interpreter tend to panic on breakpoints.

If this flag is set, then the driver forces the optimization level to 0, issuing a warning message if the command line requested a higher optimization level.

I don't understand exactly how this works. But if this flag is set *and* another condition is met, then ghc/Main.hs will alter the DynFlags so that all the hostFullWays are asked for. It is set only for the interpreter.

This flag is also special for the interpreter: if a message about a module needs to be shown, do we know anything special about where the module came from? The Boolean argument is a recomp flag.

This flag says whether the back end supports Haskell Program Coverage (HPC). If not, the compiler driver will ignore the `-fhpc` option (and will issue a warning message if it is used).

This flag says whether the back end supports foreign import of C functions. (Supports means "does not barf on," so -fno-code supports foreign C imports.)

This flag says whether the back end supports foreign export of Haskell functions to C.

When using this back end, it may be necessary or advisable to pass some `-D` options to a C compiler. This (defunctionalized) function produces those options, if any. An IO action may be necessary in order to interrogate external tools about what version they are, for example.

The function's type is Logger -> DynFlags -> IO [String]

This field is usually defaulted.

This (defunctionalized) function generates code and writes it to a file. The type of the function is

   Logger
-> DynFlags
-> Module -- ^ module being compiled
-> ModLocation
-> FilePath -- ^ Where to write output
-> Set UnitId -- ^ dependencies
-> Stream IO RawCmmGroup a -- results from `StgToCmm`
-> IO a

This (defunctionalized) function tells the compiler driver what else has to be run after code output. The type of the function is

   TPipelineClass TPhase m
=> PipeEnv
-> HscEnv
-> Maybe ModLocation
-> FilePath
-> m (Maybe FilePath)

Somewhere in the compiler driver, when compiling Haskell source (as opposed to a boot file or a sig file), it needs to know what to do with the code that the backendCodeOutput writes to a file. This Phase value gives instructions like "run the C compiler", "run the assembler," or "run the LLVM Optimizer."

Name of the back end, if any. Used to migrate legacy clients of the GHC API. Code within the GHC source tree should not refer to a back end's name.

When foreign C import or export is invalid, the carried value enumerates the valid back ends.

When foreign C import or export is invalid, the carried value enumerates the valid back ends.

Is the platform supported by the Native Code Generator?